Occasional Paper Series - No 19 / September 2021 Growth-at-risk and macroprudential policy design

Page created by Cody Harper
 
CONTINUE READING
Occasional Paper Series
No 19 / September 2021

Growth-at-risk and macroprudential
policy design
by
Javier Suarez
Contents

Executive summary 3

1 Introduction 5

2 A basic formulation of the empirical GaR approach 9

3 Social preferences and the optimal policy rule 11

 3.1 The optimal policy rule 12

 3.2 Graphical illustration 14

 3.3 Optimal target gap property 15

4 A framework for policy assessment? 17

5 Extensions 19

 5.1 Policy variable with non-linear effects 19

 5.2 Risk variable with non-linear effects 20

 5.3 A vector of policy variables 20

 5.4 Optimal policy mixes 21

 5.5 Intermediate objectives and targeted policy tools 22

 5.6 A discrete policy variable 23

6 Further discussion 25

 6.1 What if the policy variable involves no trade-off? 25

 6.2 Country heterogeneity 25

 6.3 Interaction with other policies 26

 6.4 Reformulation in terms of growth-given-stress 27

7 Concluding remarks 28

References 29

Occasional Paper Series No 19 / September 2021
Contents 1
Appendix. Microfoundations of the GaR-based welfare criterion 31

 A.1 CARA preferences and normally distributed growth rates 31

 A.2 Modelling GaR vs. growth volatility and departing from normality 33

 A.3 Reformulation using a growth-given-stress criterion 34

Imprint and acknowlegements 36

Occasional Paper Series No 19 / September 2021
Contents 2
Executive summary

This paper explores a potential application of the empirical growth-at-risk (GaR) approach to the
assessment and design of macroprudential policies. In parallel to the concept of value-at-risk, the
GaR of an economy over a given horizon is a low quantile of the distribution of the (projected) GDP
growth rate over the same horizon. In contrast to the standard macroeconomic focus on the
expected value (and, perhaps, the variance) of aggregate output growth, looking at low quantiles of
such growth implies, as in risk management, a focus on the severity of potential adverse outcomes.

The recent impulse to use the concepts of GDP-at-risk (Cecchetti, 2008) and GaR (Adrian et al.,
2018) is mainly empirical. It is related to the availability of econometric techniques that extend
regression analysis (single dependent variable models, panel data models and vector auto-
regressive models) to quantiles and their use in macroprudential applications by a growing number
of authors. Conceptually, relative to other indicators of financial stability, GaR features the
advantage of having an explicit and intuitive statistical interpretation and being measured in the
same units as GDP growth, the most universal summary indicator of an economy’s overall
performance. However, existing empirical efforts still lack a clear fit with an explicit policy design
problem of the type considered for other macroeconomic policies. This paper aims to fill this gap.

The analysis relies on a simple linear representation of the empirical GaR approach in combination
with a linear-quadratic social welfare criterion that rewards expected GDP growth and penalises the
gap between expected GDP growth and GaR. Akin to the mean-variance approach in portfolio
theory, it is shown that, if the growth rate follows a normal distribution, this welfare criterion is
consistent with expected-utility maximisation under preferences for GDP levels exhibiting constant
absolute risk aversion.

The baseline formulation – which abstracts from the time dimension by considering cumulative
growth over the relevant policy horizon – focuses on the case in which macroprudential policy
design is facing a trade-off: the available policy instrument can linearly increase GaR but at the
expense of reducing expected GDP growth. Under the baseline specification, the optimal policy rule
is linear in a variable named the “risk indicator” which represents the exogenous drivers of systemic
risk. The sensitivity of the optimal policy to changes in this risk indicator is independent of the risk
preferences embedded in the welfare criterion. This sensitivity depends directly on the impact of
risk on the gap between expected growth and GaR, and inversely on the effectiveness of policy in
reducing this gap. Optimal macroprudential policy targets a gap between expected growth and GaR
which does not depend on the level of the risk indicator but on the cost-effectiveness of
macroprudential policy and the risk preference parameter.

The explored variations in the basic setup cover cases with non-linearities in the impacts of the
policy variable and the risk variable on the relevant outcomes, multiple policy variables and discrete
policy variables. An important extension shows the compatibility of the GaR framework with the
view that macroprudential policy involves various well-identified intermediate objectives, each of
which can be associated with one or a subset of targeted policy tools. Additional discussions deal
with the case of policies which seem to involve no trade-off between mean growth and GaR, the
treatment of country heterogeneity, the interaction with other policies and the possibility of
reformulating the analysis around the concept of growth-given-stress rather than GaR.

Occasional Paper Series No 19 / September 2021
Executive summary 3
Under the postulated representation of preferences, the policy design problem yields a quantitative-
based policy target and a metric for the assessment of policy stance similar to that of other
macroeconomic policies. The main challenges for the applicability of this framework are more
empirical and political than conceptual. On the empirical side, the main challenge resides in the
consistent and sufficiently precise estimation of the causal effects of risk and policy variables on the
relevant moments (mean and GaR) of the growth distribution. Properly detecting relevant non-
linearities and interactions between policies is also important. Thus, the framework will develop at
the speed with which data on the applied policies accumulate and econometric efforts succeed in
providing reliable estimates of their effects on growth outcomes.

On the political side, once data and estimation provide a reliable description of the policy trade-offs,
the main challenge will be to define society’s aversion for financial instability on which optimal
policies should be based. Additionally, given the uncertainty surrounding the relevant parameters
implied by the empirical challenges, policymakers may need to be guided on how to expand the
type of framework sketched in this paper to account for model uncertainty (that is, for the imperfect
knowledge of the specification and parameters of the relevant quantile regressions) and the
potential policy mistakes that could stem from this uncertainty.

JEL Classification: G01, G20, G28

Keywords: macroprudential policy, policy stance, growth-at-risk, quantile regressions

Occasional Paper Series No 19 / September 2021
Executive summary 4
1 Introduction

This paper is motivated by the growing attention paid to growth-at-risk (GaR) in the assessment of
macroprudential policies. The concept arose as a natural extension to the assessment of systemic
risk of value-at-risk – a popular risk management concept. In risk management, the value-at-risk of
a given portfolio position is the critical level of the estimated distribution of possible losses over a
reference horizon that realised losses will not exceed with a high probability (such as 95% or 99%),
known as the confidence level of the assessment. Estimating the value-at-risk allows the portfolio
holder to assess, for example, the capital position that would be needed to absorb the potential
losses over the reference horizon with this confidence level of probability. From a statistical
viewpoint, the value-at-risk of a portfolio is just the estimate of a low quantile (5% or 1% in the
above examples) of the distribution of the value of the portfolio by the end of the reference horizon.

In parallel to the concept of value-at-risk, the GaR of an economy over a given horizon is a low
quantile of the distribution of the (projected) GDP growth rate over the same horizon. In other
words, the growth rate at which the probability of the realised growth rate falling below it equals a
low benchmark level such as 10% or 5%. 1 In contrast to the standard macroeconomic focus on the
expected value (and, perhaps, the variance) of aggregate output growth, looking at low quantiles of
such growth implies, as in risk management, a focus on the severity of potential adverse outcomes.
In addition to measuring this severity, the approach can provide information on the variables that
determine the probability or severity of bad outcomes, including policy variables that might then be
used to influence or “manage” the aggregate risk.

The rising popularity of GaR in financial stability and macroprudential policy assessments is driven
by demand and supply factors. From the demand side, macroprudential policy assessment and
design is in need of a quantitative framework that provides a baseline for policymakers’
discussions, decision-making and communication with the public similar to those provided by
standard macroeconomic models, targets and indicators in the fields of monetary or fiscal policy.
For macroprudential policies, the multiplicity of tools, the multidimensional nature (and still vaguely
defined concept) of systemic risk, data limitations and the relatively short historical experience with
the use of most policy tools pose significant challenges for the development of such a framework.
As a result, macroprudential policy is largely assessed and developed following a piecemeal
approach (that is, splitting the task by sector, tool, risk or detected vulnerability, or by a combination
of approaches) and relying on expert judgement for the qualitative integration of the pieces. While
the aim is to cover the whole financial system, the resulting assessment is often less complete,
integrated, systematic and quantitative than in other policy fields.

From the supply side, the impulse to use the concepts of GDP-at-risk (Cecchetti, 2008) and GaR
(Adrian et al., 2018) is mainly empirical. It is related to the availability of econometric techniques

1
 The use of lower implied confidence levels (90%, 95% in the examples above) in GaR than in value-at-risk is partly related
 to the fact that GDP is not observed at frequencies that allow for an accurate estimation of extremely low quantiles.

Occasional Paper Series No 19 / September 2021
Introduction 5
that extend regression analysis (single dependent variable models, panel data models and vector
auto-regressive models) to quantiles and their use in macroprudential applications by a growing
number of authors. Quantile-regression techniques allow the focus to be shifted from modelling the
conditional mean of the dependent variable to modelling the conditional quantiles, and thus the
whole conditional distribution of the dependent variables.

Quantile regressions allowed Cecchetti and Li (2008) to use the concept of GDP-at-risk as an
empirically-viable summary measure of the impact of asset price booms on financial stability. This
approach was further developed and promoted by the influential paper published by Adrian et al.
(2019), which shows that the lower quantiles of the distribution of the US GDP growth rate fluctuate
more and are more influenced by financial conditions than the upper quantiles, thus supporting the
focus of macroprudential surveillance and policies on the lower quantiles. Adrian et al. (2018)
documented the “term structure” of GaR and suggested the existence of an intertemporal trade-off
whereby some policies might improve GaR at medium and long horizons but at the cost of
damaging GaR (or expected growth) at shorter horizons.

Other contributions following a quantile-regression approach to the analysis of growth
vulnerabilities and their relationship with financial conditions and macroprudential policies include
Caldera-Sánchez and Röhn (2016), De Nicolo and Lucchetta (2017), Prasad et al. (2019), Arbatli-
Saxegaard et al. (2020), Chavleishvili et al. (2020), Duprey and Ueberfeldt (2020), Franta and
Gambacorta (2020), Figueres and Jarociński (2020), Galán (2021) and Aikman et al. (2021). The
empirical approach to GaR has also been embraced in part of the work undertaken by the Expert
Group on “Macroprudential Stance – Phase II” of the European Systemic Risk Board (ESRB).

Most of this empirical work puts the emphasis on the capacity of financial variables to forecast low
quantiles of GDP growth (and not necessarily high quantiles in a symmetric manner), thus
suggesting a connection between financial factors or financial stability indicators and downside risk
to output growth. 2 Other contributions focus on the impact of macroprudential policies on GaR. For
instance, Duprey and Ueberfeldt (2020), with data from Canada, find that the growth of credit to
households contributes to tail risk and that the tightening of macroprudential policy (as captured by
a qualitative index of policy actions) reduces tail risk but possibly at the cost of reducing mean GDP
growth. 3

Franta and Gambacorta (2020) also find positive financial stability implications of policy actions
relating to the tightening of loan-to-value ratios and provisioning of loan losses in a sample of 52
countries but they find no evidence of a cost in terms of mean growth outcomes. Likewise, Aikman
et al. (2021) find that higher bank capitalisation improves GaR over a three-year horizon without

2
 However, Plagborg-Møller et al. (2020) question the short-term forecasting capacity gained by considering variables such
 as the national financial conditions index (NFCI) in the prediction of GDP growth moments other than the conditional mean,
 while Brownlees and Souza (2021) challenge the out-of-sample short-term forecasting performance of quantile regressions
 relative to standard volatility models such as GARCH.
3
 The empirical analysis in Duprey and Ueberfeldt (2020) is complemented by a simple macroeconomic model that provides
 a microfoundation for the trade-off between mean growth and tail risk faced by macroprudential policy.

Occasional Paper Series No 19 / September 2021
Introduction 6
significantly reducing mean growth. In contrast, the results in Galán (2021) are consistent with the
view that the positive effect of macroprudential policy on tail outcomes over the medium term might
come at the expense of a negative effect of tightening actions on mean growth in the short term. 4

Conceptually, relative to other indicators of financial stability, GaR features the advantage of having
an explicit and intuitive statistical interpretation and being measured in the same units as GDP
growth, the most universal summary indicator of an economy’s overall performance. Hence,
quantitative contributions relating to the concept of GaR are followed with great interest (and some
scepticism too) by the institutions involved in the assessment and design of macroprudential
policies. Many see in the GaR approach a promising step in the development of an integrated
quantitative framework for macroprudential policy assessment and design. 5 However, as further
discussed in Cecchetti and Suarez (2021), existing empirical efforts still lack a clear fit with an
explicit policy design problem of the type considered for other macroeconomic policies (e.g. in the
derivation of an optimal monetary policy rule).

This paper aims to fill this gap by digging into the potential application of the empirical GaR
approach to the design and assessment of macroprudential policies. Relying on a stylised
representation of the type of equations that the quantile-regression approach may deliver, the
paper studies how macroprudential policy could be designed and evaluated using a linear-quadratic
social welfare criterion that rewards expected GDP growth and penalises the gap between
expected GDP growth and GaR. It shows that, in specific environments, this welfare criterion can
be microfounded as consistent with expected-utility maximisation under risk-averse preferences for
GDP levels. The paper characterises the properties of optimal macroprudential policy rules in the
basic setup and a number of relevant extensions. Implications are drawn on the possibility of
assessing macroprudential policy stance with a metric emanating from the estimated equations of
the empirical GaR approach.

The baseline formulation – which abstracts from the time dimension by considering cumulative
growth over the relevant policy horizon – focuses on the case in which macroprudential policy
design is facing a trade-off: the available policy instrument can linearly increase GaR but at the
expense of reducing expected GDP growth (e.g. because the tightening of some prudential
requirement reduces, within the policy horizon, medium-term vulnerabilities but has a contractive

4
 All this evidence must be taken with caution because of the hard-to-treat endogeneity of macroprudential policy actions, the
 short time span over which authorities have applied active macroprudential policies so far and the measurement difficulties
 associated with the diversity of macroprudential tools (whose activation or deactivation in many cases can only be captured
 as changes in binary variables or counting processes).
5
 Some skeptics have doubts about the feasibility and/or desirability of such an integrated approach. They think that the
 multidimensionality of macroprudential policy cannot be subsumed by looking at a single aggregate indicator such as GaR.
 Instead, a policymaker in this field might keep track of a welfare criterion that directly combines (intermediate) objectives
 along the many dimensions of systemic risk and takes into account how (potentially interacting) policies affect all such
 (intermediate) objectives. In the EU, Recommendation ESRB/2013/1 establishes five intermediate objectives for
 macroprudential policy.

Occasional Paper Series No 19 / September 2021
Introduction 7
short-term impact on economic activity). 6 Under the baseline formulation, the optimal policy rule is
linear in a variable named the “risk indicator” which represents the exogenous drivers of systemic
risk. The sensitivity of the optimal policy to changes in this risk indicator is independent of the risk
preferences embedded in the welfare criterion. This sensitivity depends directly on the impact of
risk on the gap between expected growth and GaR, and inversely on the effectiveness of policy in
reducing this gap. Optimal macroprudential policy targets a gap between expected growth and GaR
which does not depend on the level of the risk indicator but on the cost-effectiveness of
macroprudential policy and the risk preference parameter.

The explored variations in the basic setup cover cases with non-linearities in the impacts of the
policy variable and the risk variable on the relevant outcomes, multiple policy variables and discrete
policy variables. An important extension shows the compatibility of the GaR framework (and the
main insights from the basic formulation) with the view that macroprudential policy involves various
well-identified intermediate objectives, each of which can be associated with one or a subset of
targeted policy tools. Additional discussions deal with the case of policies which seem to involve no
trade-off between mean growth and GaR, the treatment of country heterogeneity, the interaction
with other policies and the possibility of reformulating the analysis around the concept of growth-
given-stress rather than GaR.

The paper is structured as follows. Section 2 provides a basic linear formulation of the type found in
the empirical GaR approach. Section 3 develops the welfare criterion used for optimal policy design
under this formulation, and derives and establishes the properties of the optimal macroprudential
policy rule. Section 4 discusses the implications of the results for the assessment of
macroprudential policy stance (that is, how the estimates associated with the empirical counterpart
of the model equations could help inform about the stance of macroprudential policy). Section 5
develops several extensions of the basic setup, generalising its results for a variety of empirically-
and policy-relevant cases, including the situation in which macroprudential policy comprises several
intermediate objectives which can be addressed with targeted tools. Section 6 contains further
discussion of the proposed analytical framework and results. Section 7 concludes the paper. The
Appendix contains the microfoundations of the GaR-based welfare criterion used in the design of
the optimal policies and discusses the extent to which, when departing from normality, a focus on
the low tail of the GDP growth distribution over a given horizon could have advantages over an
alternative focus only on the conditional mean and conditional variance of the growth distribution.

6
 The description of the macroprudential policy problem as one in which the policymaker faces a frontier in the mean growth
 vs. GaR (or tail risk) space can also be explicitly found in existing literature, including Aikman et al. (2018) and Duprey and
 Ueberfeldt (2020). However, these contributions do not elaborate on the social welfare criterion that is relevant in such a
 setting or on the properties of the implied optimal policies. Previously, Poloz (2014) referred in purely narrative/graphical
 terms to a policy frontier between financial stability risk and inflation target risk.

Occasional Paper Series No 19 / September 2021
Introduction 8
2 A basic formulation of the empirical GaR
approach

A quantile-regression approach can deliver equations for arbitrary quantiles of GDP growth over
relevant horizons. Let us consider a stylised representation of this approach that consists of two
estimated equations: one for the mean (or perhaps the median) of the (cumulative) GDP growth
over the policy horizon, denoted by �, and another for a relevant low quantile of the (cumulative)
GDP growth over the same horizon, . 7 The subscript in identifies the threshold probability (or
confidence level) at which GaR is measured. By definition, satisfies:

 Pr( ≤ ) = . (1)

This means that the probability of experiencing growth rates that are lower than over the relevant
horizon is just . The confidence level can be thought to be 5% or 10% so that reflects how bad
growth may be under adverse circumstances typically associated with systemic distress.

To start with, we consider the simple case in which the quantile-regression approach delivers
conditional forecast equations for GaR and expected growth � of the form:

 = + + , (2)

and

 � = + + , (3)

where is a unidimensional risk indicator or exogenous driver of systemic risk (e.g. a driver of
excessive credit growth or any other factor that potentially contributes to the accumulation of
financial imbalances) and is a unidimensional macroprudential policy variable (e.g. a bank capital-
based measure such as the countercyclical capital buffer (CCyB) in Basel III). 8 Let us further
assume that the endogeneity of has been treated well enough to allow for and to be
interpreted as the causal impact of variations in on GaR and expected growth, respectively.

We also assume that:

 < min{0, } and < 0 < . (4)

7
 As mentioned in the introduction, this formulation abstracts from the exact shape of the path followed by GDP growth within
 the policy horizon by focusing on the cumulative growth over the whole horizon. Practical applications may consider
 quarterly variations within a multi-quarter horizon, making it possible to capture the policy trade-off referred to below as one
 in which policy can improve GaR in a distant quarter only at the cost of reducing mean growth in a closer quarter.
8
 An advanced reader might easily extend some of the derivations and claims contained in this note to the cases in which 
 and are vectors of risk drivers and policy variables, respectively. See Section 5 for extensions of the basic formulation
 that deal with multiple policy variables. Section 6.3 considers the interaction with policies other than macroprudential
 policies.

Occasional Paper Series No 19 / September 2021
A basic formulation of the empirical GaR approach 9
In other words, risk driver has a negative impact on GaR and a less negative (or even positive)
impact on expected growth, while policy variable has a positive impact on GaR but a negative
impact on expected GDP growth. 9 The last properties imply that the policy measured by involves
a trade-off. 10 For example, if measures a driver of excessive credit growth and is the CCyB rate,
a trade-off can arise because increasing the CCyB rate reduces the final systemic risk implied by,
for instance, a credit boom (e.g. the probability and implications of an abrupt reversal) but at the
same time has a contractive impact on aggregate demand and, hence, on the central outlook. 11

Finally, we assume that the variation ranges of and , together with the values of intercepts and
 , guarantee < � over the relevant range (otherwise the linearity in (2) and (3) might lead to � <
 which would not make sense for low values of ).

9
 The linear specification implies that monotonically affects and . What really matters for the validity of the analysis
 below is that this is locally true over the relevant range of variation in . Otherwise the specification could be modified by
 redefining as a suitable non-monotonic transformation of the policy variable.
10
 Empirical findings in Adrian et al. (2018), Duprey and Ueberfeldt (2020) and Galán (2021) are consistent with the existence
 of this trade-off but findings of other authors are not (e.g. because they imply = 0). Section 6.1 discusses the case in
 which policy involves, or seems to involve, no trade-off.
11
 The risk indicator should be thought of as an exogenous driver of risk and not the final systemic risk faced by the
 economy. Systemic risk would be the result of the interaction of the risk driver and policy put in place to mitigate or
 counter its impact on tail outcomes. Thus, in the linear formulation above, systemic risk would be proportional to + 
 rather than directly and solely . In a recursive context, could also be interpreted as the predetermined value (at the time
 of deciding on policy) of a risk indicator whose evolution over the policy horizon is affected by .

Occasional Paper Series No 19 / September 2021
A basic formulation of the empirical GaR approach 10
3 Social preferences and the optimal policy
rule

To further illustrate the policy trade-offs that derive from the empirical GaR formulation, let us
suppose that the policymaker has preferences that can be represented by the social welfare
function:
 1
 = � − ( � − )2 , (5)
 2

where > 0 measures the aversion for financial instability, which is here proxied by the magnitude
of the quadratic deviations of GaR with respect to expected growth.

As shown in Section A.1 of the Appendix, in the particular case in which GDP growth follows a
normal distribution, the welfare criterion in (5) can be justified as consistent with the maximisation of
the expected utility of a representative risk-averse agent whose utility depends on GDP levels.
Specifically, if the agent has preferences for GDP levels that exhibit a constant absolute risk
aversion (CARA) coefficient , then (5) provides an exact representation of such preferences under
a value of which is directly proportional to .

Of course, if is normally distributed, social preferences and the policy problem could have also
been formulated in the usual mean-variance terms of portfolio theory, with an equation describing
the dependence of the standard deviation of the growth rate on and replacing (2) (see
Section A.2 of the Appendix for details). What this means is that the true advantages of adopting a
GaR approach (instead of a mean-variance approach) in the formulation of the macroprudential
policy problem must derive from that fact that in reality: (i) the conditional distribution of is not
Gaussian, and (ii) as documented in recent empirical work, the financial factors and policy tools on
which macroprudential policy focuses affect the conditional low quantiles of the true growth
distribution in a stronger and more clearly identifiable manner than its conditional variance. From
this perspective, an advantage of the quantile-regression approach to the modelling of the quantile
 is that it does not require the assumption of a specific distribution for the conditional quantile. In
other words, nothing prevents the estimated version of equations (2) and (3) from capturing
features such as the potential left skewness of the true conditional distribution of the GDP growth
rate. 12

Beyond the exact expected-utility microfoundations of the specific normal case, the welfare criterion
in (5) could also be defended in heuristic or axiomatic terms as the representation of the

12
 A draft policy report by Cecchetti and Suarez (2021) explores the accuracy with which the welfare measure in (5)
 approximates the underlying expected utility of a representative agent in a number of empirically motivated examples in
 which (i) the GDP growth rate is not normally distributed, and (ii) preferences on GDP levels do not exhibit CARA but rather
 constant relative risk aversion (CRRA), as typically assumed in other applications in economics and finance. For realistic
 levels of variability of cumulative GDP growth over three-year periods, the accuracy of the metric provided by (5) is very
 good.

Occasional Paper Series No 19 / September 2021
Social preferences and the optimal policy rule 11
preferences of a policymaker that is facing a trade-off between improving mean outcomes and
reducing the severity very bad outcomes. An interesting feature of (5) from such a perspective is
that the dislike for “very bad outcomes” is proportional to the square of the distance between the
bad outcomes and the mean outcomes �, where the latter would play the role of a reference
level (or status quo point) similar to those emphasised in some non-expected-utility formulations of
agents’ preferences for risk. Specifically, from the perspective of prospect theory (Kahneman and
Tversky, 1979), the coefficient in (5) could be interpreted as capturing loss aversion rather than
risk aversion. 13

3.1 The optimal policy rule
An optimal macroprudential policy that is conditional on a risk level would thus maximise given
 . That is, it would be characterised by the policy rule:

 ( ) = argmax ( , ), (6)
 
where ( , ) describes as a function of the risk indicator and the policy variable after taking
into account (2) and (3).

If the optimal policy is interior, it must solve the following first order condition (FOC):
 � � 
 − ( � − ) � − � = 0, (7)
 
which uses the chain rule in (5). From (2) and (3), this FOC can be written as:

 − ( + + − − − )( − ) = 0. (8)

Solving for leads to the macroprudential policy rule:

 ( ) = 0 + 1 , (9)

with
 − 
 0 = + (10)
 − ( − )2

and
 − 
 1 = . (11)
 − 

13
 The asymmetric focus on low tail losses can also be related to Fishburn (1977), who explores preferences in which the
 decision maker is averse to obtaining below-target payoffs. Kilian and Manganelli (2008) analyse the decision problem of a
 central banker using that approach. In a similar vein, Svensson (2003) considers a monetary policy problem under
 preferences that asymmetrically penalise extreme events.

Occasional Paper Series No 19 / September 2021
Social preferences and the optimal policy rule 12
Under our assumptions, the intercept of the policy rule 0 can, in principle, have any sign since it is
the sum of a first term which will most typically be positive (specifically if − > 0) and a second
term which is negative (since < 0). However, 0 is intuitively increasing in the policymaker’s
preference for financial stability (since the absolute size of the negative term declines with ) and
also increasing in the difference − > 0, which measures the effectiveness of the policy variable
in reducing the gap between expected growth and GaR, � − . 14

Interestingly, the parameter 1 which measures the responsiveness of the optimal policy to
variations in risk indicator is positive and independent of the preference parameter . So in this
setup, policymakers with different preferences for financial stability would differ in the level at which
they use macroprudential policy but not in the extent to which they modify their policies in response
to changes in the risk assessment. This optimal policy responsiveness is directly proportional to the
impact of risk on the gap between expected growth and GaR ( − , which is positive under (4))
and inversely proportional to the effectiveness of policy in reducing the gap ( − , which is also
positive under (4)).

While the empirical GaR approach (namely, estimating equations (2) and (3)) does not, per se,
allow the policy parameter to be estimated, it might allow a direct estimate to be made of the
optimal policy responsiveness parameter 1 and its components − and − . It might also
allow the optimal policy rule to be represented for different illustrative values of the preference
parameter . 15

14
 For instance, in the polar case in which the policymaker has absolute preference for financial stability ( → ∞), the intercept
 would become just ( − )/( − ) and lead to a solution with � = (which, although unrealistic in practice, is
 mathematically feasible given the linearity of (2) and (3) in and ). In the other polar scenario with no preference for
 financial stability ( → 0), we would have 0 → −∞, implying that the policymaker would choose the lowest possible value
 of , since under the linear specification of (3) this is the way to maximise expected growth (albeit at the cost of minimising
 GaR).
15
 The use of the conditional “might” is a reminder of the importance of relying on estimates of parameters and that reflect
 the causal impact of policy on growth outcomes and not just partial correlations between historical realisations of policy and
 outcomes.

Occasional Paper Series No 19 / September 2021
Social preferences and the optimal policy rule 13
3.2 Graphical illustration
Further understanding of the interaction between the policy trade-offs implied by (2) and (3) and the
preferences reflected in (5) can be obtained by depicting the frontier of pairs of and � that can be
reached, for a given value of risk variable , by varying the policy variable . Mathematically, this
conditional policy frontier (for a given ) is defined by the line:

 � = � − � + � − � + , (12)
 
which is downward sloping in ( , �) space. Figure 1 depicts the policy frontier for a given value of
 . The point ( ( , 0), �( , 0)) corresponds to the case in which = 0. Intuitively choosing > 0
allows higher values of to be obtained, but at the cost of lowering �. 16

Figure 1
Graphical illustration of the policy design problem

16
 Under the assumed linearity, there is nothing special about = 0 but in specific applications the policy variable could be
 normalised so that it means something, e.g. the historical mean or “normal stance” of the corresponding policy (then < 0
 would represent a stance that is looser than normal and > 0 a stance that is tighter than normal). For some policy
 instruments there may be a natural lower bound to , e.g. a CCyB rate under Basel III cannot be negative (although in
 practice there are instances of capital forbearance that might be similar to having < 0 for this instrument). Explicit
 consideration of these bounds would raise complications regarding occasionally binding constraints that are familiar in
 other contexts.

Occasional Paper Series No 19 / September 2021
Social preferences and the optimal policy rule 14
The preferences in (5) describe a map of indifference curves in ( , �) space that are convex
parabolas which reach their minima on the ray = �. The map makes economic sense to the left
of this ray. Intuitively, for > 0, on each indifference curve any decline in should be
compensated with an increase in � in order to keep the welfare level unchanged. Moreover, for a
given decline in , the required compensating increase in � increases with the distance from the
ray � = . This explains why the FOC (7) includes the term ( � − ), which accounts for the
marginal cost of financial instability.

Optimal policy ( ) is the choice of that leads to maximum welfare on the corresponding
conditional policy frontier. In other words, ( ) is the policy level that leads to the point where the
conditional policy frontier is tangent to the map of indifference curves. From the determinants of the
slopes of these curves it follows that, all other things being equal, a policymaker with a stronger
preference for financial stability will choose combinations of ( , �) on the frontier that involve a
lower � and a higher , that is, a lower gap between expected growth and GaR.

3.3 Optimal target gap property
What happens when the risk indicator moves? From (12), changes in the risk indicator cause a
parallel shift in the policy frontier (just another implication of the linear formulation). In the most
plausible situation, in which risk does not increase expected growth too much (formally, when <
 / , where / is positive under (4)), a rise in shifts the policy frontier down, necessarily
worsening the terms of choice for the policymaker. In the alternative scenario, where > / ,
risk has such a strong effect on expected growth that it moves the policy frontier up.

In both cases, however, the optimal policy rule (9) implies that an increase in risk will lead to a
tightening of policy decision , indicating that the fall in that would occur if policy were not
adjusted is at least partly offset by increasing . When risk has a positive marginal impact on
expected growth ( > 0), the optimal policy response will diminish the raw positive effect of risk
indicator on mean growth �. When risk has a negative marginal impact on expected growth ( <
0), the optimal policy response will still aim to offset its even more negative effect on by lowering
mean growth � beyond the implications of the raw negative effect of risk indicator .

Mathematically, the final impact of changes in risk indicator on and � can be seen by
substituting the optimal policy rule ( ) in (2) and (3), which leads to:

 − 
 = ( + 0 ) + ( + 1 ) = ( + 0 ) + , (13)
 − 

and
 − 
 � = ( + 0 ) + ( + 1 ) = ( + 0 ) + . (14)
 − 

Interestingly, the coefficient of risk indicator is identical in these two equations, which implies that
the optimal policy rule would keep the gap between expected growth and GaR constant:
 1 (− ) 1 1
 � − = ( − ) + ( − ) 0 = = (15)
 − 1+ /(− )

Occasional Paper Series No 19 / September 2021
Social preferences and the optimal policy rule 15
Notice that this target gap is positive under (4) since < 0. The target gap decreases in the
preference for financial stability and increases in the marginal growth-gap rate of transformation
implied by the policy frontier, (− )/( − ), which can be rewritten as 1/[1 + /(− )] to better
visualise its negative dependence with respect to the marginal cost-effectiveness of the policy
variable: the ratio of the -quantile-improving effect > 0 to the mean-reducing effect − < .

In fact, this “constant target gap” property can be directly obtained from the FOC in (7), which can
be rearranged as:
 �
 
 1 1 1
 � − = − 
 � 
 = . (16)
 − 1+ /(− )
 
Graphically, this implies that changes in risk indicator and the optimal policy response under ( )
describe a linear expansion path in ( , �) with a slope equal to one. Thus, starting from the optimal
policy identified in Figure 1 for a particular risk level , changes in will lead to combinations ( , �)
on the line with slope one that goes through that point.

More specifically, when risk does not increase expected growth too much (that is, in the case of
 < / described above), the coefficient of in the reduced-form equations (13) and (14) is
negative. Thus, when the risk indicator increases and policy responds optimally, GaR and expected
growth deteriorate by the same amount, keeping the gap between expected growth and GaR, � −
 constant. This is the case depicted in Figure 1, where the policy frontier under ′ > lies on the
left of the policy frontier for .

Otherwise (when > / ), the coefficient of in (13) and (14) is positive, and rises in and the
optimal policy response lead GaR and expected growth to improve by the same amount, but once
again keeping � − constant.

An important corollary to these findings is that, under the specified preferences, macroprudential
policy should not target a constant GaR or keep GaR above a certain lower bound, but should allow
the GaR target to comove (actually by the same amount) with the expected growth estimate. In
other words, these derivations suggest that the � − gap is a more useful indicator of stance than
each of its components separately.

Occasional Paper Series No 19 / September 2021
Social preferences and the optimal policy rule 16
4 A framework for policy assessment?

A tentative list of policy-relevant outputs that this approach can deliver is set out below.

1. Estimating (2) and (3) allows us to positively describe the direct impact of risk and policy on
 GaR and expected growth, as well as the policy trade-offs involved.

2. The policy trade-offs can be further illustrated using a policy frontier as shown in Figure 1.

 (a) If it is evaluated at the historical mean value of , this frontier could be called the mean
 policy frontier. If a practical application involves a discrete , then a reference value
 could be selected to represent a “normal” situation.

 (b) Under the linear specification, the conditional policy frontier is just a parallel shift of the
 mean (or “normal”) policy frontier. The relative position of the conditional frontier relative
 to the historical mean (or “normal”) frontier may indicate whether the economy is facing a
 state of above-normal or below-normal risk exposure.

3. If social preferences (or the preferences of the policymaker) can be described with a mean
 growth versus GaR welfare criterion as in (5), then the following points should be taken into
 consideration.

 (a) The optimal policy responsiveness to the risk indicator can be measured using 1 =
 − 
 , as in (11). This measure is independent of parameter that describes the
 − 

 policymaker’s preference for financial stability.

 (b) If the policymaker follows the optimal policy rule, it will implicitly target a constant gap
 between expected growth and GaR, as in (16). The optimal policy gap will be decreasing
 in the preference parameter and in the cost-effectiveness ratio of the policy tool
 /(− ). From this gap, and the estimates of the empirical GaR model, the implicit
 preference parameter could be recovered (inferred) from the condition:
 1 1
 = . (17)
 �− 1+ /(− )

 (c) Conditional on a reference value of , the optimal policy rule can be fully described using
 (6). Graphically, it can be described using the expansion path previously illustrated in
 Figure 1.

 (d) Conditional on a reference value of and an assessment of risk , a graphical
 counterpart of the optimal policy choice can be described by depicting the conditional
 policy frontier and the point on it that is associated with the optimal policy (given by the
 intersection between the policy frontier and the expansion path).

 (e) Conditional on an assessment of risk , a policy stance could be deemed inefficient if it
 leads to points sufficiently far away from the policy frontier. However, when the policy
 variable is unidimensional (as in all the derivations above), all choices of are

Occasional Paper Series No 19 / September 2021
A framework for policy assessment? 17
“efficient”, so the concept of inefficiency is only useful when there are two or more policy
 variables (as in some of the extensions discussed below).

 (f) Conditional on the reference value of and an assessment of risk , a policy stance
 could be deemed suboptimal if it is sufficiently far away from the expansion path. This
 corresponds to an excessive distance between and ( ) or, in terms of outcomes, a
 � − gap that is sufficiently far away from its target. Thus, policy would be too tight if 
 is sufficiently higher than ( ) and equivalently if the � − gap is well below the target.
 Conversely, policy would be too loose if is sufficiently lower than ( ) and equivalently
 if the � − gap is well above target.

Occasional Paper Series No 19 / September 2021
A framework for policy assessment? 18
5 Extensions

This section considers several specific extensions of the basic formulation. It demonstrates the
capacity of the framework to accommodate multiple variations and checks the robustness of the
properties of optimal macroprudential policies in each of them.

5.1 Policy variable with non-linear effects
A particularly relevant non-linearity may be related to the diminishing effectiveness of the policy
variable (or variables, if there are several) in improving GaR. Another interesting case may emerge
if the impact of the policy variable on expected growth is marginally increasing. Let us therefore
consider a generalised version of (2) and (3) in which

 = + + Γ ( ), (18)

and

 � = + + Γ( ), (19)

where the functions Γ ( ) and Γ( ) satisfy Γ ′ < 0 < Γ ′ and Γ ′′ < Γ ′′ < 0. In this case, the FOC
solved by an interior optimal policy can be written as:

 Γ ′ ( ) − [ �( , ) − ( , )][Γ ′ ( ) − Γ ′ ( )] = 0, (20)

where the dependence of � and on and has been made explicit to emphasise the type of
non-linear equation that would have to be solved to find the optimal policy rule ( ).

By rearranging (20), we obtain an expression for the gap associated with the optimal policy that is
very similar to (16):
 1 1
 �( , ) − ( , ) = . (21)
 1+Γ′ ( )/(−Γ′ ( ))

However, in this case the target gap is not invariant to risk indicator . If increases, all other things
being equal, the left hand side of (21) increases, calling for an offsetting increase in . But the right
hand side of (21) is now increasing in because, intuitively, the policy trade-off measured by the
marginal cost-effectiveness of the policy (here Γ ′ ( )/(−Γ ′ ( ))) worsens at higher levels of . This
implies that optimal policy ( ) in this case grows less than linearly with and the optimal gap
increases with . In other words, as risk deteriorates, the policymaker would accommodate the
diminishing cost-effectiveness of the policy tool by widening the targeted gap between expected
growth and GaR.

Occasional Paper Series No 19 / September 2021
Extensions 19
5.2 Risk variable with non-linear effects
Let us now consider a situation in which risk variable has a non-linear impact on expected growth
and GaR, captured by the following modification of (2) and (3):

 = + ( ) + , (22)

and

 � = + ( ) + , (23)

where ( ) and ( ) are functions satisfying ′ ( ) − ′ ( ) > 0 and ′′ ( ) − ′′ ( ) > 0. In other
words, increases in widen the gap between expected GDP and GaR at an increasing rate (e.g. by
making the financial system more and more likely to reach a tipping point of full meltdown). In this
case, the FOC of the welfare maximisation problem becomes:

 − [ �( , ) − ( , )]( − ) = 0, (24)

which implicitly defines the policy rule ( ). In this case, the FOC, and consequently the policy rule,
are only non-linear because of the non-linear effect of on �( , ) − ( , ). Rearranging this to
solve for the optimal gap, we obtain:
 1 1
 �( , ) − ( , ) = , (25)
 1+ /(− )

where the right hand side is invariant to , thus implying the same gap as when had a linear
impact on and �. Interestingly, in this case, when the risk indicator increases, the left hand side
increases more than proportionally to , also calling for a more than proportional increase in the
policy variable. In this setup, the policy response to rises in risk should be increasingly aggressive
as risk increases. 17

5.3 A vector of policy variables
Let us consider an extended version of (2) and (3) with different continuous policy variables 
with = 1,2, . . . linearly affecting and � with coefficients and , respectively. We assume
that these coefficients satisfy < 0 < as in (4) and further assume that the variables are scaled
so that = 0 is the lowest bound applicable to all of them. In this linear world, as seen by exploring
the relevant first order conditions, there will generally be one variable dominating the others in the
maximisation of . This most efficient or preferred policy tool ∗ would be the one featuring the
lowest value of what was referred to as the marginal growth-gap rate of transformation in the single
policy variable benchmark:

17
 The opposite situation, in which optimal policy is increasingly less aggressive as increases. emerges if ′′ ( ) − ′′ ( ) <
 0.

Occasional Paper Series No 19 / September 2021
Extensions 20
�
 
 1
 � 
 = > 0, (26)
 − 1+ /(− )
 
that is, the policy tool with the best marginal cost-effectiveness as measured by /(− ).
Intuitively, when /(− ) is higher, the same reduction in the gap between expected growth and
GaR can be achieved at a lower cost in terms of expected growth. For the most efficient tool, the
optimal value of ∗ would be the value that satisfies the counterpart of equation (7). The associated
policy rule would be the same as in (9) with 0 and 1 particularised to the preferred tool ∗ . All
elements in Figure 1 remain valid if the policy frontiers are also particularised to those obtained
using the preferred tool.

All the other policy variables should remain at their lowest bound of zero. In terms of Figure 1, using
an inferior tool would imply moving over policy “frontiers” that also go through the point
( ( , 0), �( , 0)) but with steeper slopes, confirming that such a tool would only be able to increase
 by causing larger declines in �. Conditional on using a less cost-effective tool, equation (16)
would imply that the target gap should be larger, thus accommodating the harder trade-off faced
along the corresponding policy frontier.

5.4 Optimal policy mixes
The optimality of using non-trivial combinations of tools in macroprudential policy would only
emerge in the event of departures from linearity. For example, optimal policies could be obtained
that involve using several tools at the same time if the effectiveness of each policy tool in reducing
 ′
GaR is marginally decreasing, for instance, as given by some functions Γ ( ) with Γ > 0 and
 ′′
Γ < 0, or if there are complementarities between tools under a general quasi-concave function
 
Γ ( 1 , 2 , . . . ) that replaces the terms Σ =1 in the extended version of (2).

In such a non-linear world, all policy variables activated at a strictly positive level at the optimum
would satisfy a properly modified version of (7) and, consequently, (16), implying:
 �
 
 1 1 1
 � − = − � 
 = . (27)
 − 1+ Γ /(− )
 
Thus, optimal policy mixes would feature equalisation of the marginal cost-effectiveness ratios,
 Γ 
 /(− ), across all the activated policy tools. The optimal gap between expected GDP growth and
 
GaR would be decreasing in both the common cost-effectiveness ratio and the aversion to financial
instability.

In terms of interactions between tools, the optimal gap may no longer be constant since the
compound effectiveness of a given policy mix may depend on the intensity with which policies are
activated. For example, if a rise in the risk variable calls for a more intensive use of two
complementary policies that jointly exhibit decreasing returns to intensity (akin to when
complementary inputs are combined in a production function with decreasing returns to scale), then
the optimal policy will accommodate (as in the case of the single policy variable with decreasing
marginal effectiveness discussed above) the decreasing effectiveness by tolerating a larger gap
when the risk is high than when the risk is low.

Occasional Paper Series No 19 / September 2021
Extensions 21
5.5 Intermediate objectives and targeted policy tools
Current practice of macroprudential policy largely involves a piecemeal approach. Authorities
around the globe, as well as research in the field of macroprudential policy, often address the
design and assessment of macroprudential policy by splitting it into separate silos. As in
microprudential regulations, these silos are commonly determined by the nature of the underlying
source of systemic risk (e.g. liquidity vs. solvency risk) or by the sector that originates, transmits or
suffers the risk (e.g. banks vs. non-banks, commercial vs. residential real estate, etc.). The
resulting silos are typically associated with one or several dedicated policy tools (e.g. “capital-based
tools for the banking sector”, “liquidity-based tools for the investment management sector” or
“borrower-based tools for residential real estate risk”). The practical attractiveness of the piecemeal
approach stems from the difficulties of integrating under a common general equilibrium perspective
and with a common ultimate goal the multiple dimensions of systemic risk, the multiple potential
factors contributing to financial stability (or the lack of it) and the multiple policy tools available to
address these dimensions and factors. The purpose of this subsection is to show that such an
approach is not incompatible with the analytical framework and empirical efforts associated with the
GaR approach. In fact, the latter can contribute to integrating, adding up or at least bringing under a
common umbrella sectoral macroprudential policies that might otherwise be difficult to relate to
each other when trying to obtain an overall notion of macroprudential policy stance.

The illustration shown below of the capability of the GaR approach to integrate prior piecemeal
approaches into macroprudential policy design relies on simplifying assumptions that are directed
at making the problem analytically tractable. In the spirit of keeping things simple and close to some
of the extensions described above, let us assume that macroprudential policy involves different
dimensions, = 1,2, . . and that each dimension can now be associated with an intermediate
objective and a targeted policy tool . 18

Let us further assume that intermediate objectives can be represented as linear functions of their
targeted tools:

 = 0 + 1 , (28)

where 0 is an autonomous component of the intermediate objective and 1 > 0 measures the
marginal impact of the targeted policy variable on the intermediate objective. 19 The baseline
equations (2) and (3) could then be reformulated as follows:

 = + Γ ( 1 , 2 , . . . ), (29)

and

18
 Advanced readers may further expand the proposed setup to accommodate additional generalisations of the problem.
19
 Without loss of generality, we impose a sign convention for and such that increasing is good for financial stability (i.e.
 it increases ) and increasing the policy variable is good for the intermediate objective (i.e. it increases ).

Occasional Paper Series No 19 / September 2021
Extensions 22
 
 � = + Σ =1 , (30)

where Γ is an increasing and strictly concave function of the vector of intermediate objectives and
 < 0 so that, as in the baseline setup, macroprudential policies involve a trade-off: increasing
policy improves the intermediate objective but at the cost of reducing mean growth at the
margin. 20

As in the non-linear world described in the above extension on optimal policy mixes, all targeted
policy variables that are activated at a strictly positive level at the optimum would satisfy a modified
version of (7) and, consequently, (16), implying:
 �
 
 1 1 1
 � − = − � 
 = . (31)
 − 1+ Γ /(− )
 
Thus, the optimal vector of targeted policies would again feature equalisation of the marginal
growth-gap rates of transformation and hence equalisation of the marginal cost-effectiveness ratios
across all activated policy tools. Moreover, the optimal gap between expected GDP growth and
GaR would be decreasing in both the aversion to financial instability and the common ratio. The
cost-effectiveness ratio of targeted policy is the ratio between the marginal effectiveness of the
 Γ
policy, that is, its marginal capability to improve GaR by affecting intermediate objective ( )
 
and the marginal cost of the policy in terms of mean growth (− ).

Depending on the degree to which intermediate objectives may feature complementarity in their
compound impact on GaR, as captured by the cross derivatives of function Γ , the setup with
multiple intermediate objectives might imply increasing or decreasing the target gap as well as
varying the optimal policy mix in response to changes, for instance, in the autonomous component
of one of the intermediate objectives. For example, in the simple case in which Γ is additively
separable across intermediate objectives (so that they do not directly interact in affecting GaR), the
policy response to a deterioration in the autonomous component of one objective would be the
 Γ
tightening of policy across all intermediate objectives (so that declines across all policy
 ′
 ′
dimensions and the equality in (31) is restored at a higher target gap).

5.6 A discrete policy variable
Intuitively, if the policy variable is discrete and yet enters the problem as assumed in (2) and (3),
the left hand side of the FOC in (7) must be replaced by its finite differences counterpart and its
sign checked to discover whether there are gains from increasing (or keeping on increasing) the
variable or, conversely, whether there could be gains from reducing it.

20
 Notice that, for purposes of simplicity and in contrast to the baseline specification, (29) and (30) do not explicitly contain any
 risk variable . However, it would be trivial to introduce one or a vector of them affecting and � linearly as in the baseline
 model. We could also consider risk variables that affect the autonomous component 0 of each intermediate objective.

Occasional Paper Series No 19 / September 2021
Extensions 23
You can also read